openclaw - 💡(How to fix) Fix Regression: Plugin commands fail to intercept agent run on Telegram (v2026.5.4 still affected) [1 comments, 2 participants]

Description

Despite the fix reported in #77384 (shipped in v2026.5.3), custom plugin commands registered via api.registerCommand() are still failing to intercept agent runs in specific scenarios.

The LLM (agent) continues to respond to registered slash commands instead of letting the local plugin handler execute and terminate the turn.

Environment

• OpenClaw Version: v2026.5.4 (325df3e)
• OS: macOS 26.4.1 (arm64)
• Channel: Telegram
• Channel Config: commands.native: false (Custom commands are handled as text routing rather than native Telegram UI commands)

Steps to Reproduce

1. Register a custom command in a plugin using api.registerCommand({ name: 'test', ... }).
2. Ensure commands.native: false is set for the Telegram channel.
3. Send /test to the bot.

Expected Behavior

The plugin handler should execute, and the turn should end (unless continueAgent: true is explicitly returned).

Actual Behavior

The agent run is triggered, and the LLM responds conversationally (e.g., explaining what the command might mean or just chatting), ignoring the local interceptor.

Additional Context

• This is a regression that appeared in v2026.5.2.
• The re-ordering of loadCommandHandlers() mentioned in #77384 does not seem to cover the text-routing path for Telegram when native commands are disabled.
• Rolling back to v2026.4.23 immediately restores the correct interception behavior.

Reference: Replaces and reopens the case from #77384.